Research map / Continual learning / Continual fine-tuning of LLMs
Affordance-First Decomposition for Continual Learning in Video-Language Understanding
Mengzhu xu, Hanzhi Liu, Ningkang Peng, Qianyu Chen, Canran Xiao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add affordance tokens (vs direct frame tokens to LLM) (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 48.7 -> 51.6 (-2.9 when removed) | Improved |
| add query-conditioned router (vs uniform adapter mixing) (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 49.8 -> 51.6 (-1.8 when removed) | Improved |
| add conflict-triggered LoRA rank growth (vs fixed rank 8) (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 50.5 -> 51.6 (-1.1 when removed) | Improved |
| add question-only replay distillation (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 50.2 -> 51.6 (-1.4 when removed) | Improved |
| add ASR weak-alignment term in affordance loss (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 50.8 -> 51.6 (-0.8 when removed) | Improved |
| add teacher-consistency KL term in affordance loss (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 50.6 -> 51.6 (-1.0 when removed) | Improved |
| use soft Top-L mixture (L=8) instead of hard Top-L=1 (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 50.7 -> 51.6 (-0.9 when removed) | Improved |
| use larger memory budgets (BQ=8k, BA=1k vs 2k, 256) (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 51.0 -> 51.6 (-0.6 when removed) | Improved |
| add affordance tokens (ablation) | AFD · Domain-incremental VideoQA (6 datasets) | forgetting: -3.3 -> -1.8 (+1.5 when removed) | Improved |
| add affordance tokens (ablation) | AFD · ViLCo-Bench MQ R@1@0.5 | R@1@0.5 (%): 28.0 -> 29.6 (-1.6 when removed) | Improved |
| AFD vs DAM, best prior continual method | AFD · Domain-incremental VideoQA (6 datasets) | average accuracy (%): 50.2 -> 51.6 (+1.4) | Improved |
| AFD vs DAM on time-incremental iVQA | AFD · Time-incremental iVQA (4 slices) | average accuracy (%): 38.1 -> 39.5 (+1.4) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Bisecle — Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 2025) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- $\infty$-Video: A Training-Free Approach to Long Video Understanding via Continuous-Time Memory Consolidation (ICML 2025 · similarity 0.89)
- On the Consistency of Video Large Language Models in Temporal Comprehension (CVPR 2025 · similarity 0.88)
- Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization (NeurIPS 2025 · similarity 0.88)
- Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 2025 · similarity 0.87)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.87)
- ViLCo-Bench: VIdeo Language COntinual learning Benchmark (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
